char-fork/llama.cpp

git clone https://git.t4t.associates/char-fork/llama.cpp

Yaniss Amazouzmodel: add NVIDIA Nemotron-3-Puzzle-75B-A9B (NemotronHPuzzle) support (#25444)c61b98b87

char/aliceai-t5-moe
brancheschar/aliceai-t5-moeall refs →
7.8 KiB194 linesraw
1#include "models.h"
2
3void llama_model_deepseek::load_arch_hparams(llama_model_loader & ml) {
4    ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
5    ml.get_key(LLM_KV_LEADING_DENSE_BLOCK_COUNT,   hparams.n_layer_dense_lead, false);
6    ml.get_key_or_arr(LLM_KV_EXPERT_FEED_FORWARD_LENGTH, hparams.n_ff_exp_arr, hparams.n_layer_all);
7    ml.get_key(LLM_KV_EXPERT_SHARED_COUNT,         hparams.n_expert_shared);
8    ml.get_key(LLM_KV_EXPERT_WEIGHTS_SCALE,        hparams.expert_weights_scale, false);
9
10    switch (hparams.n_ff_exp()) {
11        case 1408: type = LLM_TYPE_16B; break;
12        case 1792: type = LLM_TYPE_20B; break;
13        default: type = LLM_TYPE_UNKNOWN;
14    }
15}
16
17void llama_model_deepseek::load_arch_tensors(llama_model_loader &) {
18    LLAMA_LOAD_LOCALS;
19    const int64_t n_expert_shared = hparams.n_expert_shared;
20
21
22    const int64_t n_ff_exp        = hparams.n_ff_exp();
23
24    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
25
26    // output
27    output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
28    // try to load output.weight, if not found, use token_embd (tied embeddings)
29    output      = create_tensor(tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, TENSOR_NOT_REQUIRED);
30    if (!output) {
31        output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED);
32    }
33
34    for (int i = 0; i < n_layer; ++i) {
35        auto & layer = layers[i];
36
37        layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
38
39        create_tensor_qkv(layer, i, n_embd, n_embd, n_embd_gqa, n_embd_gqa, 0);
40        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}, 0);
41        layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
42
43        if (i < (int) hparams.n_layer_dense_lead) {
44            layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd,   n_ff}, 0);
45            layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd}, 0);
46            layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff}, 0);
47        } else {
48            layer.ffn_gate_inp = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP, "weight", i), {n_embd, n_expert}, 0);
49
50            if (n_expert == 0) {
51                throw std::runtime_error("n_expert must be > 0");
52            }
53            if (n_expert_used == 0) {
54                throw std::runtime_error("n_expert_used must be > 0");
55            }
56
57            // MoE branch
58            layer.ffn_gate_exps = create_tensor(tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), {  n_embd, n_ff_exp, n_expert}, 0);
59            layer.ffn_down_exps = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {n_ff_exp,   n_embd, n_expert}, 0);
60            layer.ffn_up_exps   = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS,   "weight", i), {  n_embd, n_ff_exp, n_expert}, 0);
61
62            // Shared expert branch
63            layer.ffn_gate_shexp = create_tensor(tn(LLM_TENSOR_FFN_GATE_SHEXP, "weight", i), {n_embd, n_ff_exp * n_expert_shared}, 0);
64            layer.ffn_down_shexp = create_tensor(tn(LLM_TENSOR_FFN_DOWN_SHEXP, "weight", i), {        n_ff_exp * n_expert_shared, n_embd}, 0);
65            layer.ffn_up_shexp   = create_tensor(tn(LLM_TENSOR_FFN_UP_SHEXP,   "weight", i), {n_embd, n_ff_exp * n_expert_shared}, 0);
66        }
67    }
68}
69
70std::unique_ptr<llm_graph_context> llama_model_deepseek::build_arch_graph(const llm_graph_params & params) const {
71    return std::make_unique<graph>(*this, params);
72}
73
74llama_model_deepseek::graph::graph(const llama_model & model, const llm_graph_params & params) :
75    llm_graph_context(params) {
76    const int64_t n_embd_head = hparams.n_embd_head_v();
77
78    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
79    GGML_ASSERT(n_embd_head == n_rot);
80
81    ggml_tensor * cur;
82    ggml_tensor * inpL;
83
84    inpL = build_inp_embd(model.tok_embd);
85
86    // inp_pos - contains the positions
87    ggml_tensor * inp_pos = build_inp_pos();
88
89    auto * inp_attn = build_attn_inp_kv();
90
91    const float kq_scale =
92        hparams.f_attention_scale == 0.0f ? 1.0f / sqrtf(float(n_embd_head)) : hparams.f_attention_scale;
93
94    ggml_tensor * inp_out_ids = build_inp_out_ids();
95
96    for (int il = 0; il < n_layer; ++il) {
97        ggml_tensor * inpSA = inpL;
98
99        // norm
100        cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);
101        cb(cur, "attn_norm", il);
102
103        // self-attention
104        {
105            // rope freq factors for llama3; may return nullptr for llama2 and other models
106            ggml_tensor * rope_factors = model.get_rope_factors(cparams, il);
107
108            // compute Q and K and RoPE them
109            auto [Qcur, Kcur, Vcur] = build_qkv(model.layers[il], cur,
110                    n_embd_head, n_head, n_head_kv, il);
111
112            Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, rope_factors, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
113                                 ext_factor, attn_factor, beta_fast, beta_slow);
114
115            Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, rope_factors, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
116                                 ext_factor, attn_factor, beta_fast, beta_slow);
117
118            cb(Qcur, "Qcur", il);
119            cb(Kcur, "Kcur", il);
120            cb(Vcur, "Vcur", il);
121
122            cur = build_attn(inp_attn,
123                    model.layers[il].wo, model.layers[il].wo_b, model.layers[il].wo_s,
124                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);
125        }
126        if (il == n_layer - 1 && inp_out_ids) {
127            cur   = ggml_get_rows(ctx0, cur, inp_out_ids);
128            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
129        }
130        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
131        cb(ffn_inp, "ffn_inp", il);
132
133        cur = build_norm(ffn_inp, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il);
134        cb(cur, "ffn_norm", il);
135
136        if ((uint32_t) il < hparams.n_layer_dense_lead) {
137            cur = build_ffn(cur,
138                    model.layers[il].ffn_up, NULL, NULL,
139                    model.layers[il].ffn_gate, NULL, NULL,
140                    model.layers[il].ffn_down, NULL, NULL,
141                    NULL, LLM_FFN_SILU, LLM_FFN_PAR, il);
142            cb(cur, "ffn_out", il);
143        } else {
144            // MoE branch
145            ggml_tensor * moe_out = build_moe_ffn(cur,
146                model.layers[il].ffn_gate_inp,
147                model.layers[il].ffn_up_exps,
148                model.layers[il].ffn_gate_exps,
149                model.layers[il].ffn_down_exps,
150                nullptr,
151                n_expert, n_expert_used,
152                LLM_FFN_SILU, false,
153                hparams.expert_weights_scale,
154                LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX,
155                il);
156            cb(moe_out, "ffn_moe_out", il);
157
158            // FFN shared expert
159            {
160                ggml_tensor * ffn_shexp =
161                    build_ffn(cur,
162                        model.layers[il].ffn_up_shexp, NULL, NULL,
163                        model.layers[il].ffn_gate_shexp, NULL, NULL,
164                        model.layers[il].ffn_down_shexp, NULL, NULL,
165                        NULL, LLM_FFN_SILU, LLM_FFN_PAR, il);
166                cb(ffn_shexp, "ffn_shexp", il);
167
168                cur = ggml_add(ctx0, moe_out, ffn_shexp);
169                cb(cur, "ffn_out", il);
170            }
171        }
172        cur = ggml_add(ctx0, cur, ffn_inp);
173
174        cur = build_cvec(cur, il);
175        cb(cur, "l_out", il);
176
177        // input for next layer
178        inpL = cur;
179    }
180    cur = inpL;
181
182    cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1);
183
184    cb(cur, "result_norm", -1);
185    res->t_embd = cur;
186
187    // lm_head
188    cur = build_lora_mm(model.output, cur, model.output_s);
189
190    cb(cur, "result_output", -1);
191    res->t_logits = cur;
192
193    ggml_build_forward_expand(gf, cur);
194}